Quotient space model of hierarchical query-by-humming system

In this paper, we put ideas of granular computing into the application of query-by-humming. Leading by quotient space theory of problem solving, we present a quotient space model of hierarchical query-by-humming system, which do search in two stages. The first is searching in the quotient space to obtain high recall rate and a narrowed possible region with accepted speed. The second step is doing accurate but relatively slow search in the small original space to achieve high precision. In addition, we improve the previous dynamic programming techniques for melody matching, which tolerate the input error more reasonably. We also compare our hierarchical method with the previous flat note-based algorithm and frame-based algorithm on MIDI-encoded database of Chinese music and obtain promising results both in efficiency and accuracy.